Contract type : Fixed-term contract

Level of qualifications required : Graduate degree or equivalent

Fonction : PhD Position

The job will be performed at INRIA in Sierra team. The PhD will be funded by the ERC Starting Grant REAL.

Context of REAL: Current machine learning is not suitable to deal with the new scenario, both from a theoretical and a practical viewpoint: the lack of cost-effectiveness of the algorithms impacts directly the eco- nomic/energetic costs of large scale machine learning, making it barely affordable by universities or research institutes, and the lack of reliability of the predictions affects critically the safety of the systems where machine learning is employed.

To deal with the challenges posed by the new scenario, REAL will lay the foundations of a solid theoreti- cal and algorithmic framework for reliable and cost-effective large scale machine learning on modern computational architectures. In particular, REAL will extend the classical supervised learning framework to provide algorithms with two additional guarantees: (a) the predictions will be reliable, i.e., endowed with explicit bounds on their uncertainty guaranteed by the theory (e.g., confidence intervals in case of regression); (b) the algorithms will be cost-effective, i.e., they will be naturally adaptive to the new architectures and will provably achieve the desired reliability and accuracy level, by using minimum possible computational resources.

The internship is in the context on the recently awarded ERC Starting Grant REAL. The whole goal of the research project is to help machine learning to overcome its current limitations and become an solid tool for science and engineering, from a practical and theoretical viewpoint.

In the last decade, machine learning (ML) has become a fundamental tool with a growing impact in many disciplines, from science to industry. However, nowadays, the scenario is changing: data are exponentially growing compared to the computational resources (post Moore’s law era), and ML algorithms are becoming crucial building blocks in complex systems for decision making, engineering, and science.

Standard ML techniques learn real valued functions. However nowadays many machine learning problems require to learn functions with structured output, e.g., automatic text translation (string -> string), image captioning (image -> string), speech recognition, prediction on graphs, learning on manifolds, learning to rank documents, protein folding, etc. In this context current research [1] focuses on developing ad hoc algorithms for each structured problem, often without theoretical guarantees. 

Recently, [2,4] proposed a unifying theoretical framework to address general structured prediction problems by introducing a novel learning strategy with strong statistical guarantees.  In particular the derived methods are expressed in terms of an optimization problem in the output space.  However, the topology and the inner structure of the data is crucial to devise proper learning strategies. Relevant data come from physical process or, more generally, processes with a recursive structure. A first step to leverage the internal structure of the process generating the data for proper structured prediction has been done in [5], where data constituted by a collection of loosely interacting parts has been considered.

The goal of this PhD is to extend the results in [5] by designing novel machine learning techniques for structured prediction that are able to benefit from the intrinsic topology of the data, beyond the horizontal kind of structures analyzed in [5] and fully exploiting the recursive (fixed-point) structure of the generating process. It would be interesting to further develop optimization strategies to efficiently solve structured prediction problems in practice and deal with large scale datasets. The work is at the intersection of algorithms, statistics and optimization, and may focus primarily on any these three aspects depending on the candidate.

Students interested in this project should contact Alessandro Rudi to discuss further.

[1] Gükhan H. Bakir, Thomas Hofmann, Bernhard Schölkopf, Alexander J. Smola, Ben Taskar, and S. V. N. Vishwanathan. 2007. Predicting Structured Data (Neural Information Processing). The MIT Press.

[2] Ciliberto, C., Rudi, A. and Rosasco, L. A Consistent Regularization Approach for Structured Prediction. 

In Advances in Neural Information Processing Systems 2016. (pp. 4412-4420). Barcelona

[3] https://www.inria.fr/en/alessandro-rudis-erc-starting-grant-re-inventing-machine-learning

[4] A General Framework for Consistent Structured Prediction with Implicit Loss Embeddings

Carlo Ciliberto, Lorenzo Rosasco, Alessandro Rudi; 21(98):1−67, 2020.

[5] Ciliberto, Carlo, Francis Bach, and Alessandro Rudi. "Localized structured prediction." Advances in Neural Information Processing Systems. 2019

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking and flexible organization of working hours (after 12 months)
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities
• Access to vocational training
• Social security coverage